Multi-View Attention Network for Visual Dialog

Visual dialog is a challenging vision-language task in which series of questions visually grounded by given image are answered. To resolve the visual task, high-level understanding various multimodal inputs (e.g., question, history, and image) required. Specifically, it necessary for an agent to (1) determine semantic intent question (2) align question-relevant textual contents among heterogeneous modality inputs. In this paper, we propose Multi-View Attention Network (MVAN), leverages multiple views about based on attention mechanisms. MVAN effectively captures information from history with two complementary modules (i.e., Topic Aggregation Context Matching), builds representations through sequential alignment processes Modality Alignment). Experimental results VisDial v1.0 dataset show effectiveness our proposed model, outperforms previous state-of-the-art methods under both single model ensemble settings.